Graph attention convolution network for power flow calculation considering grid uncertainty

• A novel neural network model for real-time power flow calculation. • Fully considers both external and internal uncertainties of the grid. • Achieves higher computational accuracy and robustness. • Respects the physical properties of the power grid. • Effectively captures the underlying correlations within the data. With the increasing penetration of renewable energy sources and the growing complexity of power system structures, the grid is increasingly impacted by both external and internal uncertainties. In this context, probabilistic power flow models based on artificial intelligence need to possess stronger learning capabilities and robustness to handle the dynamic changes in the grid. To this end, this paper proposes a novel graph attention convolution model for real-time power flow calculation. Specifically, first, a multi-order hybrid adjacency matrix is constructed based on the impedance distance of the grid to fully reflect the underlying relationships between nodes. Second, a new correlation attention mechanism is developed to explore the correlations within the data. Meanwhile, inductive learning is used during node sampling to more efficiently utilize global information. Furthermore, the learning ability of the model is further enhanced by the combination of stacked dilated convolution layers and a fully connected layer. Finally, the model’s performance is tested on three different IEEE cases. The results show that, compared to the state-of-the-art methods, the proposed model not only achieves higher accuracy in predicting bus voltages and branch power but also demonstrates good robustness against uncertain topologies.